[PATCH v2 0/5] arm64: advertise availability of CRC and crypto instructions

On 20 January 2014 19:17, Nicolas Pitre [off-list ref] wrote:

On Mon, 20 Jan 2014, Ard Biesheuvel wrote:

quoted

On 20 January 2014 18:44, Nicolas Pitre [off-list ref] wrote:

quoted

On Mon, 20 Jan 2014, Catalin Marinas wrote:

quoted

On Mon, Dec 23, 2013 at 02:06:27PM +0000, Ard Biesheuvel wrote:

quoted

This series is a followup to the patch that was recently merged by Catalin that
allocates hwcaps bits for CRC and Crypto Extensions instructions so userland can
discover whether the current CPU has any of those capabilities.

Patch #1 enables ARM support for the ELF_HWCAP2/AT_HWCAP2 ELF auxv entry that
was recently added to the kernel and glibc (2.18). It extends the feature bit
space to 64 bits (on 32-bit architectures)

Patch #2 adds generic support for ELF_HWCAP2/AT_HWCAP2 to the 32-bit ELF compat
mode for 64-bit architectures.

Patch #3 adds support for ELF_HWCAP2/AT_HWCAP2 to arm64's 32-bit ELF compat mode

Patch #4 allocates the HWCAP2 bits in the arch/arm tree. This is necessary
because 32-bit ARM binaries can execute both under ARM and under arm64 kernels,
so there should be agreement about the meaning of feature bits, even if the ARM
kernel has no support yet for ARMv8 32-bit only hardware (such as ARMv8-R).

It looks a bit strange to start filling HWCAP2 before HWCAP is full but
I guess we want to preserve some future extensions in HWCAP for older
glibc.

How could older glibc possibly care about future extensions?

Calling getauxval(AT_HWCAP) on an outdated libc.so will give you the
whole value, not just the bits whose meaning was known to glibc at the
time.
So if desired, a program can interpret AT_HWCAP itself.

AT_HWCAP2 is completely new, so you won't be able to retrieve it using
getauxval() on an older libc.

I agree.  And I don't dispute the bit placement choice either.

Still, an old glibc calling getauxval(AT_HWCAP) should already be
prepared to receive and rightfully ignore those bits it didn't know the
meaning of at the time.  So "preserving some future extensions in HWCAP
for older glibc" as a justification makes little sense to me... unless
I'm missing something?

Even if applications interpret those bits themselves, supposing they
still need to be linked against an old glibc, then why would
yet-to-be-defined future extensions be more important to be signaled
using the lower 32 bits than the extensions you propose?  That is what I
don't get.

In the general case, you are quite right.

In this particular case, the extensions for which I am adding the
feature bits are not supported on any hardware currently known or
supported by the ARM port. At this time, the only known CPUs
supporting these extensions are ARMv8 CPUs executing in 32-bit
compatibility mode (i.e., ARMv8 defines instructions for the 32-bit
execution state using previously unallocated opcodes)

So the only reason (currently) for adding these feature bits to the
ARM port is to align with the ARMv8 32-bit compat mode so that a
32-bit userland requires no knowledge about whether its 32-bit
execution environment is hosted by an ARM or an arm64 kernel. In the
future, ARMv8 32-bit only CPUs may well turn up that support these
extensions as well.

-- 
Ard.